Anthropic has announced the release of Claude Opus 5, its latest AI model, which is now available across all platforms.
The company states that Opus 5 delivers notable improvements in efficiency and cost-effectiveness over its predecessor, Opus 4.8, while maintaining the same pricing structure.
Go deeper with GlobalData
Access deeper industry intelligence
Experience unmatched clarity with a single platform that combines unique data, AI, and human expertise.
The updated model is designed for everyday use and has become the default option on Claude Max, as well as the most advanced model available on Claude Pro.
According to Anthropic, Opus 5 is particularly efficient on software engineering and knowledge work tasks, achieving leading results on benchmarks including Frontier-Bench and GDPval-AA. However, it continues to trail Mythos 5 on certain cybersecurity evaluations.
Opus 5’s pricing remains unchanged from the previous version, at $5 per million input tokens and $25 per million output tokens. It is also available in a Fast mode which operates at approximately 2.5 times the standard speed and is offered at twice the base rate.
Developers can access the new model through the Claude API using the claude-opus-5 endpoint.
Anthropic reports that Opus 5 outperforms earlier models on benchmarks assessing software engineering, problem-solving, and business automation. For instance, on the Frontier-Bench v0.1 evaluation, Opus 5 achieves more than double the performance of Opus 4.8 at a reduced cost per task.
On CursorBench 3.2 in maximum effort mode, the model’s score is within 0.5% of Fable 5’s peak, at half the cost. Other assessments cited include the ARC-AGI 3, where Opus 5 scores three times higher than the next-best model, and the Zapier AutomationBench, with a pass rate around 1.5 times greater than alternatives at the same cost per task.
On OSWorld 2.0, it surpasses Fable 5’s best result at just over one-third of the cost.
Cursor co-founder Sualeh Asif said: “Claude Opus 5 delivers near Fable 5 intelligence at Opus speed and cost. On CursorBench it’s just under Fable 5 and has many of the same behaviours.”
Anthropic notes specific gains in scientific research-related tasks, with Opus 5 reportedly outperforming Opus 4.8 across life sciences evaluations, such as structural biology and organic chemistry.
The company highlights a 10.2% point improvement in molecular structure inference and a 7.7 point increase in protein-related tasks compared to the previous model.
Opus 5 is described as being more capable at verifying outcomes and repeating tasks until successful.
Anthropic shared several practical examples from evaluations and early-access testers, including the reconstruction of a machine part from pixel data and identifying the root cause of a bug in open-source software.
In another case, an engineer reportedly used Opus 5 to build a market data feed for a new exchange in a single session, which earlier models could not accomplish.
Regarding alignment and safety, Anthropic’s internal behavioural audit found Opus 5 to exhibit higher adherence to Claude’s Constitution and lower incidents of deceptive behaviour than previous models.
The company states Opus 5 remains behind Mythos 5 in both biological research and offensive cybersecurity, refraining from advancing in riskier, dual-use capabilities.
During testing such as OSS-Fuzz, Opus 5 reportedly identified vulnerabilities with results similar to Mythos 5 but was less capable when tasked with developing exploits.
Opus 5’s safeguards permit its use in cybersecurity and biology, subject to certain restrictions. According to Anthropic, cyber classifiers in Opus 5 are less restrictive than those on Fable 5, allowing for vulnerability identification in source code while blocking some higher-risk activities.
With the launch of Opus 5, Anthropic is also releasing two beta updates. These include mid-conversation tool changes for developers using the Claude Platform and automatic fallbacks on the API, which allow requests blocked by safety classifiers to be redirected to alternative models.
Anthropic confirmed there are no data retention requirements for general access to the Claude Opus 5 model, consistent with earlier releases in the Opus family.
Recently, Anthropic announced plans to use up to 2GW of AMD Instinct MI450 Series GPUs in AMD Helios rack-scale systems under a partnership with AMD. In addition to the infrastructure agreement, AMD has committed to a strategic equity investment of up to $5bn in Anthropic.
